As of September 1, 2026, supported image, video, and audio formats produced by Claude carry a verifiable provenance signature, and text carries a watermark. For a Quebec SMB, the question is no longer "can we tell?" but "what do we do with this in our day-to-day operations?"
The short answer: yes, it is now possible to trace a file produced by an AI back to its origin. Anthropic announced on September 1, 2026, that supported image, video, and audio formats generated by Claude through the code execution tool carry Content Credentials to the C2PA standard when retrieved through the Files API, and that text produced by Claude Fable 5.1 and Claude Mythos 5.1 carries a text watermark. No changes are required on the application side: the marking is automatic. What it does not do: it does not create a legal obligation to label your content, and it does not replace an internal AI usage policy.
What is a "Content Credential" C2PA, in practical terms?
It is an identity card attached to the file itself. The standard comes from the Coalition for Content Provenance and Authenticity (C2PA), a project of the Joint Development Foundation, an American 501(c)(6) non-profit organization that develops open technical standards for content provenance (source: c2pa.org). The coalition was founded in February 2021 by Adobe, Arm, BBC, Intel, Microsoft, and Truepic; its specification is currently at version 2.4 [publication date to be confirmed: the spec page does not display it]. Worth noting: it is the subject of an international standardization project (ISO/CD 22144, "Authenticity of information," still under development), so it is not yet a published ISO standard. The principle: instead of guessing after the fact whether an image was generated, you read the declared provenance directly from the file.
For an SMB, the value is not theoretical. It applies to three very ordinary situations:
- A visual produced by a supplier: knowing what was generated versus what was photographed.
- A file received by email in a sensitive matter (claim, incident, HR).
- Your own marketing content: being able to demonstrate your process if a client asks.
A key limitation: a file without a Content Credential is not proof that it is authentic, and provenance information can be lost when a file is recompressed or passes through a platform that does not preserve it. It is one more data point, not a verdict.
Does Law 25 require me to identify AI-generated content?
No. And it is important to say that clearly, because the confusion is common.
The Act respecting the protection of personal information in the private sector (RLRQ, c. P-39.1), as amended by Law 25, governs personal information, not content provenance. The closest obligation is section 12.1: when a decision is based exclusively on automated processing, the organization must inform the person concerned and, upon request, explain the information used and the main factors that led to the decision. Meaningful human involvement in the process changes the picture.
In other words: the real compliance question is not "was this text written by an AI?" but "is an AI making a decision about a person, on its own?". That is where the obligation lies.
C2PA marking remains useful for compliance in the broader sense: it documents your processes. An organization that can show how content was produced and who reviewed it is an organization that inspires confidence in clients, insurers, and regulators.
The question nobody asks before choosing a model: where does the data go?
This is the part of the September 1 announcement that nobody is talking about, yet it is the part with the greatest impact on a Quebec SMB.
In the same release notes, Anthropic specifies that Claude Fable 5.1 and Claude Mythos 5.1 require 30-day data retention and are not available in zero-data-retention mode without express authorization from Anthropic.
Choosing a model is a governance decision, not just a performance decision. A more powerful model may be incompatible with the confidentiality commitments you have made to your clients or with your personal information handling policy. This does not mean walking away from the model, but evaluating the implications upfront and formally documenting the decision.
This is exactly the kind of trade-off that the CAPE methodology (Conceive · Activate · Propel · Evolve) places at the Conceive stage: you decide on the framework before plugging in the tool, not the other way around.
Four concrete steps for an SMB, this quarter
- Take stock of your actual AI usage. Not the official usage: the real usage, across your teams. You will almost always find more than expected.
- Write a one-page usage policy. What can be submitted to an AI tool, what cannot, and who reviews before publication.
- Check the data retention terms of the model or models you use and record the answer in your register. The default configuration is not always what you assume.
- Keep a named human reviewer on all content and decisions that affect a person. It is the best safeguard, and it is also what keeps you outside the scope of section 12.1.
FAQ
Can I reliably detect AI-generated text? No, not in any general way. The watermark announced on September 1, 2026, applies to text produced by Claude Fable 5.1 and Claude Mythos 5.1: it is a vendor-specific mechanism, not a universal detector. Generic "AI detector" tools produce false positives.
Is a file without a Content Credential suspicious? No. The absence of provenance proves nothing: the vast majority of files in circulation do not carry one, and the information can be lost during processing.
Does Law 25 require me to disclose that I use AI? Not as such. The obligation under section 12.1 targets decisions based exclusively on automated processing that concern a person. A marketing communication drafted with AI assistance and reviewed by a human is not covered by that section.
Is it worth waiting for the regulatory landscape to clarify? The steps recommended above (inventory, usage policy, human review) have value regardless of how regulations evolve. They are low-cost and they reduce risk right now.
You are already using AI in your organization and want to know where you stand on governance and compliance? Reach out: we start from your actual usage, not a theoretical questionnaire.
LinkedIn Version
Do you know which AI tools your teams are actually using, and where the data they enter ends up?
On September 1, Anthropic announced two things in the same release notes. The first made the rounds online: supported image, video, and audio formats produced by Claude now carry a provenance signature to the C2PA standard, and text carries a watermark.
The second flew under the radar, and it is the one that matters for an SMB: these new models require 30-day data retention and are not available in zero-retention mode without express authorization.
Choosing an AI model is therefore not just a performance decision. It is a governance decision.
And no, Law 25 does not require you to label your AI-generated content. Section 12.1 targets something else: decisions based exclusively on automated processing about a person. The real question is not "who wrote this text" but "is a machine making a decision about someone on its own."
Three steps worth their weight in gold, regardless of where regulations land: take stock of your actual usage, write a one-page policy, keep a named human reviewer in the loop.
AI Law25 QuebecSMB DataGovernance DigitalTransformation
Sources cited
- Anthropic, Claude Platform release notes, entry of September 1, 2026 (text watermark, C2PA Content Credentials on image/video/audio via the Files API; 30-day data retention requirement for Claude Fable 5.1 and Claude Mythos 5.1): https://platform.claude.com/docs/en/release-notes/api, accessed 2026-09-03. (The former address
docs.anthropic.com/en/release-notes/apiredirects to this one: cite the canonical address.) - Coalition for Content Provenance and Authenticity (C2PA), About (project of the Joint Development Foundation, 501(c)(6) non-profit, open technical standards for content provenance): https://c2pa.org/about/, accessed 2026-09-03.
- C2PA, C2PA Specification, version 2.4 [publication date to be confirmed: the spec page does not display it]: https://spec.c2pa.org/specifications/specifications/2.4/index.html, accessed 2026-09-03.
- ISO, ISO/CD 22144, Authenticity of information: project under development at ISO/TC 171/SC 2, not published as of the consultation date, verified 2026-09-03.
- Act respecting the protection of personal information in the private sector, RLRQ c. P-39.1, s. 12.1 (decision based exclusively on automated processing): primary legislative reference, text to be attached at time of publication.